Study on AI coding agents finds gains "absorbed" by human review "bottleneck." (arstechnica.com)

A recent study by Harvard researchers Fiona Chen and James Stratton reveals that while AI coding agents can significantly boost the volume of code generated—by up to 30%—the efficiency gains are undermined by a substantial bottleneck in human code review processes. The study, which analyzed data from over 700 software development firms and 300 million work events, found that as coding practices integrated more AI tools, the workload on human reviewers increased. This led to a higher frequency of revisions and more extensive reviewer feedback, effectively negating the anticipated benefits of increased AI assistance. This finding is significant for the AI/ML community as it highlights a critical challenge in fully harnessing the power of AI tools in software development. Increased code production does not equate to improved output or reduced employment within organizations, as firms have not seen statistically significant changes in issue resolution metrics post-AI tool implementation. This suggests that while AI can automate and expedite parts of the coding process, existing human review practices and their constraints must be optimized to truly capitalize on the advancements offered by AI in coding.